Abstracts

Rationale: Continuous electroencephalography (cEEG) is vital to the detection of seizures in critically ill newborns. Despite the importance of monitoring neonates, the labor and resource intense nature prevents wide spread use. The aim of this retrospective study is to use clinical features, EEG indications, and initial EEG findings to create a prediction model that will identify neonates at highest risk for seizure.Methods: Retrospective study of patient’s less than 1 month corrected gestational age enrolled between January 1, 2011 and December 31, 2013 who had a clinically indicated cEEG (EEG > 3 hours) in the cardiac and neonatal intensive care unit at Boston Children’s Hospital. Electrographic seizures (ES) were defined as any seizure detected on cEEG whether electroclinical or electrographic only. Patient demographics including gestational age, and clinical data including EEG indication were collected. High risk characteristics such as the need for hypothermia in the setting of concern for hypoxic injury, patient’s requiring extracorporeal membrane oxygenation (ECMO), and patients that suffered cardiac arrest were also included. Description of the initial EEG background was based upon ACNS guidelines. The prediction model for detection of seizures was developed using logistic regression with group-lasso penalty. Clinical and initial EEG features were incorporated into the model.Results: 210 neonates were included, of these 158 were term. Preterm neonates were found to have an odds ratio 0.48 showing a lack of association with seizures as compared to term neonates. An event concerning for clinical seizure holds a higher odds ratio of 2.73 in comparison to 0.72 in neonates monitored with a concern for nonconvulsive seizures. Background features and probability of seizure was also evaluated with a focus on the presence or absence of a continuous waking background and the presence of excess discontinuity based on gestational age. The presence of a continuous background produced an odds ratio of 0.3 (Table 1). Two prediction models were created the first using clinical variables alone. The area under the curves (AUC)s for the Receiver Operating Characteristic (ROC) of the models with clinical variables only are 74.3% (67.6-81.0%) for the minimum classification error rate and 72.6% (66.0-79.4%) for the parsimonious model within 1 SE of the minimum as compared to 83.3% (77.9-88.7%) and 79.9% (74.3-85.8%) when EEG variables are included. The cut-off points of -0.732 and -0.898 provide sensitivity of 82.19 and 90.21, and specificity of 54.01 and 43.80 with clinical variables alone The cut-off points of -0.749 and -0.623 provide sensitivity of 90.41 and 87.67, and specificity of 68.61 and 65.69 respectively with addition of EEG variables (Figure 1).Conclusions: The inclusion of EEG background features considerably improves the overall prediction of seizure. This will allow better allocation of resources and development of individualized monitoring strategies for critically ill neonates at risk of seizure.